Refresh git metadata after git init

[srid]

Kolu now notices when a terminal directory becomes a git repo after the Code tab already showed the no-repo state. A newly initialized .git directory updates terminal metadata, which in turn refreshes the Code browser and pill-tree identity without waiting for another cwd change.

The git integration now keeps a lightweight .git entry watcher active while a terminal cwd is unresolved or non-git, then hands off to the existing .git/HEAD watcher once GitInfo resolves. The follow-up review commits make the subscription mode explicit and keep HEAD-vs-entry watcher lifecycle counts separate, so future watcher changes have clearer tests.

Coverage

• Unit coverage for external git init publishing without setCwd
• E2E coverage for Code tab leaving Not in a git repository and pill tree switching to repo/branch labels
• Scoped e2e run: just test-quick features/code-tab.feature features/git-context.feature

Closes #813.

Try it locally

nix run github:juspay/kolu/fix/git-init-refresh-code-browser

Generated by /do on Codex (model gpt-5).

[srid]

Hickey/Lowy Analysis

#	Lens	Finding	Disposition
1	Hickey	GitInfo null drives watcher mode	Fixed in this PR
2	Hickey	Watcher counts hide axis	Fixed in this PR

Hickey rationale

The initial implementation used GitInfo | null both as the current metadata value and as the watcher-mode flag. That spread the invariant across subscribeGitInfo: null meant the .git entry watcher should be active, non-null meant HEAD watching should own repo changes. The fix introduced explicit watcher mode ownership and then tightened the implementation to track actual active watcher axes, so a no-op pre-repo HEAD watcher is not mistaken for a live watcher.

The tests also had generic installs / retires counters that only measured HEAD watcher logs. The fix renamed these to HEAD-specific counters and added .git entry watcher counters, so lifecycle assertions describe the resource being measured.

Lowy rationale

No Lowy findings. The new .git entry watcher is a narrow activity boundary for repo-appearance detection, while subscribeGitInfo remains the orchestration boundary for resolving, watching, and publishing metadata. Server and client modules continue to depend only on GitInfo | null, so the repo-appearance volatility stays inside kolu-git.

[srid]

Evidence

• Started a terminal in /tmp/kolu-evidence-git-init-refresh before it was a git repository.
• Opened the Code tab and verified the no-repo state.
• Ran git init and checked out code-refresh from outside the browser session.
• Verified the Code tab refreshed to the empty-changes state and the pill tree showed kolu-evidence-git-init-refresh / code-refresh.

[srid]

/do results

Step	Status	Duration	Verification
sync	✓	0s	git fetch ok; forge=github; noGit=false
research	✓	3m 45s	Issue #813 mapped: CodeTab uses active metadata git repoRoot, pill tree derives terminalKey from git/cwd, subscribeGitInfo only rechecks git init on same-cwd setCwd; fix belongs in packages/integrations/git/src/resolve.ts with tests in packages/integrations/git/src/index.test.ts and packages/tests/features/code-tab.feature
branch	✓	10s	Created feature branch fix/git-init-refresh-code-browser from origin/master
implement	✓	2m 22s	Added shared .git entry watcher for non-repo cwd metadata, unit coverage for external git init without setCwd, and e2e coverage for Code tab plus pill-tree refresh
check	✓	25s	just check exited 0: pnpm typecheck passed across workspace; biome lint completed with warnings but no failing diagnostics
docs	✓	39s	Updated README Git/Architecture text and packages/integrations/git/README.md for the .git entry watcher; packages/surface/README.md unchanged because no surface descriptor changed
fmt	✓	12s	just fmt completed: Biome and nixpkgs-fmt reported no file changes
commit	✓	21s	Committed 19b26e3 (fix: refresh git metadata after external init) and pushed origin/fix/git-init-refresh-code-browser
hickey+lowy	✓	6m 45s	Hickey and Lowy completed. Hickey found two fix-in-PR items, addressed by commits 37b8e83 and acc31db; Lowy reported no findings. No unactioned findings remain.
police	✓	3m 8s	code-police clean after fixes: rules found no-silent-error-swallowing comment gap fixed in 5e59202; fact-check found active watcher-axis bug fixed in 748ff8e; elegance clean. Focused kolu-git unit tests passed and just check exited 0.
test	✓	39s	just test-quick features/code-tab.feature features/git-context.feature passed: 28 scenarios, 304 steps
create-pr	✓	52s	Draft PR #822 created at #822 with title/body matching scope; Hickey/Lowy analysis comment posted with two fixed Hickey findings and no Lowy findings
ci	✗	113m 11s	CI on 4935654: all expected contexts posted; only ci/e2e@aarch64-darwin remains failing after full retry and targeted retry, both with libc++abi Napi::Error/server-crash flake in .logs/4935654/e2e@aarch64-darwin.log
evidence	✓	0s	Posted ## Evidence PR comment with screenshot proving external git init refresh: #822 (comment)
Total		134m 45s

Slowest step: ci (113m 11s)

Optimization suggestions

• CI dominated this run. For the next attempt, resume from CI only and target CI_SYSTEM=aarch64-darwin just ci::e2e unless code changes again.
• The repeated Darwin failure is still the known libc++abi ... Napi::Error server-crash path; hardening that test/server crash would remove most retry time for this PR class.
• Evidence now has a direct Playwright path for this issue, so future reruns can reuse the uploaded screenshot workflow instead of manually revalidating the UI.

Workflow failed at 2026-05-04T19:48:08Z.